Project Background:
The project background in Spiking Neural Networks (SNNs) emerges from the quest to develop more biologically plausible and energy-efficient artificial intelligence models inspired by the workings of the human brain. SNNs represent a departure from traditional artificial neural networks by mimicking the spiking behavior of neurons, where information is processed and transmitted in the form of discrete spikes or action potentials. This paradigm shift offers several advantages and robustness to noisy inputs. Additionally, SNNs exhibit remarkable capabilities in tasks such as pattern recognition, sensor fusion, and neuromorphic computing, making them attractive for a wide range of applications. It involves understanding the principles of neural information processing, synaptic plasticity, and spike-timing-dependent plasticity (STDP) that underpin SNNs. Moreover, it encompasses efforts to develop novel training algorithms, neuromorphic hardware platforms, and software frameworks for simulating and implementing SNNs efficiently.
Problem Statement
- Traditional artificial neural networks lack biological plausibility, failing to capture the spiking behavior of neurons and synaptic dynamics observed in the brain.
-
Conventional neural networks require significant computational resources and energy consumption, limiting their scalability and applicability to resource-constrained devices.
-
Sparse and Asynchronous Processing aim to emulate what is observed in the brain, facilitating more efficient information representation and transmission.
-
Developing effective learning algorithms for SNNs, especially in the context of spatiotemporal spike-based learning rules, presents a significant challenge due to the complex dynamics of synaptic plasticity.
-
Translating SNNs from theoretical models to practical hardware implementations poses challenges in designing efficient neuromorphic hardware architectures capable of real-time processing and learning.
- Ensuring the robustness and generalization capabilities of SNNs across diverse tasks and datasets remains a key challenge, requiring improvements in network architectures, training methodologies, and regularization techniques.
- Scaling up SNNs to handle large-scale and complex tasks while maintaining computational efficiency and biological realism poses technical challenges in network design, training, and optimization.
Aim and Objectives
- To develop biologically-inspired Spiking Neural Networks (SNNs) for efficient and robust information processing.
-
Mimic is the spiking behavior of neurons observed in biological systems.
-
Improve computational efficiency and energy consumption compared to traditional artificial neural networks.
-
Develop effective learning algorithms for training SNNs, leveraging spatiotemporal spike-based learning rules.
-
Design efficient neuromorphic hardware architectures for real-time implementation of SNNs.
-
Enhance the robustness and generalization capabilities of SNNs across diverse tasks and datasets.
-
Scale up SNNs to handle large-scale, complex computational tasks while maintaining efficiency and biological realism.
Contributions to Spiking Neural Networks
- Advancing the understanding of neural information processing by mimicking biological spiking behavior.
-
Improving computational efficiency and energy consumption compared to traditional artificial neural networks.
-
Developing novel learning algorithms tailored to spatiotemporal spike-based learning rules.
-
Proposing efficient neuromorphic hardware architectures for real-time implementation of SNNs.
-
Enhancing the robustness and generalization capabilities of SNNs across various tasks and datasets.
-
Scaling up SNNs to handle large-scale and complex computational tasks while maintaining efficiency and biological realism.
Deep Learning Algorithms for Spiking Neural Networks
- SpikeProp
-
Temporal Contrastive Divergence (TCD)
-
Spike-based Backpropagation (SBP)
-
Spike Timing-Dependent Plasticity (STDP)
-
ReSuMe
-
Spike-timing-dependent optimization (STDO)
-
Deep Belief Spiking Network (DBSN)
-
Spike Temporal Encoding and Processing (STEP)
-
Temporal encoding and decoding machines (TEDM)
Datasets for Spiking Neural Networks
- N-MNIST
-
N-Caltech101
-
DVS Gesture
-
SpiNNaker DVS Gesture Dataset
-
HReNet Dataset
-
NeuCube Datasets
-
Silicon Retina Dataset
-
Event-based Dataset for Object Tracking (EDOT)
-
Neuromorphic Benchmark Datasets (NBD)
-
Loihi DVS Gesture Dataset
Software Tools and Technologies
Operating System: Ubuntu 18.04 LTS 64bit / Windows 10
Development Tools: Anaconda3, Spyder 5.0, Jupyter Notebook
Language Version: Python 3.9
Python Libraries:
1.Python ML Libraries:
- Scikit-Learn
- Numpy
- Pandas
- Matplotlib
- Seaborn
- Docker
- MLflow
2.Deep Learning Frameworks:
- Keras
- TensorFlow
- PyTorch
